class SHAInet::Network

• SHAInet::Network
• Reference
• Object

Defined in:

shainet/basic/exprimental.cr
shainet/basic/network_run.cr
shainet/basic/network_setup.cr

Constant Summary

CONNECTION_TYPES = ["full", "ind_to_ind", "random"]

COST_FUNCTIONS = ["mse", "c_ent"]

LAYER_TYPES = ["input", "hidden", "output"]

Log = ::Log.for(self)

Constructors

• .new

  First creates an empty shell of the entire network

Instance Method Summary

• #add_layer(l_type : Symbol | String, l_size : Int32, n_type : Symbol | String = "memory", activation_function : ActivationFunction = SHAInet.sigmoid)

  Create and populate a layer with neurons l_type is: :input, :hidden or :output l_size = how many neurons in the layer n_type = advanced option for different neuron types

• #all_neurons : Array(SHAInet::Neuron)

  General network parameters

• #all_synapses : Array(SHAInet::Synapse)

  General network parameters

• #alpha : Float64

  Parameters for Adam

• #alpha=(alpha : Float64)

  Parameters for Adam

• #b_gradient : Array(Float64)
• #beta1 : Float64
• #beta2 : Float64
• #clean_dead_neurons
• #connect_ltl(src_layer : Layer, dest_layer : Layer, connection_type : Symbol | String)

  Connect two specific layers with synapses

• #delta_max : Float64

  Parameters for Rprop

• #delta_max=(delta_max : Float64)

  Parameters for Rprop

• #delta_min : Float64

  Parameters for Rprop

• #delta_min=(delta_min : Float64)

  Parameters for Rprop

• #epsilon : Float64
• #error_signal : Array(Float64)
• #etah_minus : Float64

  Parameters for Rprop

• #etah_minus=(etah_minus : Float64)

  Parameters for Rprop

• #etah_plus : Float64

  Parameters for Rprop

• #etah_plus=(etah_plus : Float64)

  Parameters for Rprop

• #evaluate(input_data : Array(GenNum), expected_output : Array(GenNum), cost_function : CostFunction = SHAInet.quadratic_cost)

  Quantifies how good the network performed for a single input compared to the expected output This function returns the actual output and updates the error gradient for the output layer

• #evaluate_exp(input_data : Array(GenNum), expected_output : Array(GenNum), cost_function : CostFunction = SHAInet.quadratic_cost, stealth : Bool = true)
• #fully_connect

  Connect all the layers in order (input and output don't connect between themselves): input, hidden, output

• #get_cost_proc(function_name : String) : CostFunction
• #hidden_layers : Array(SHAInet::Layer)

  General network parameters

• #input_layers : Array(SHAInet::Layer)

  General network parameters

• #inspect

  Returns an unambiguous and information-rich string representation of this object, typically intended for developers.

• #learning_rate : Float64

  Parameters for SGD + Momentum

• #learning_rate=(learning_rate : Float64)

  Parameters for SGD + Momentum

• #load_from_file(file_path : String)
• #log_summary(e)
• #momentum : Float64

  Parameters for SGD + Momentum

• #momentum=(momentum : Float64)

  Parameters for SGD + Momentum

• #mse : Float64
• #output_layers : Array(SHAInet::Layer)

  General network parameters

• #prev_mse : Float64
• #randomize_all_biases
• #randomize_all_weights
• #run(input : Array(GenNum), stealth : Bool = false) : Array(Float64)

  Run an input throught the network to get an output (weights & biases do not change)

• #run_exp(input : Array(GenNum), stealth : Bool = false) : Array(Float64)
• #save_to_file(file_path : String)
• #test(test_set)

  Evaluate the network performance on a test set

• #time_step : Int32
• #total_error : Float64
• #train(data : Array(Array(Array(GenNum))) | SHAInet::TrainingData, training_type : Symbol | String, cost_function : Symbol | String | CostFunction = :mse, epochs : Int32 = 1, error_threshold : Float64 = 0.00000001, mini_batch_size : Int32 = 1, log_each : Int32 = 1000, show_slice : Bool = false, autosave : NamedTuple(freq: Int32, path: String) | Nil = nil)

  Training the model ameba:disable Metrics/CyclomaticComplexity

• #train_batch(data : Array(Array(Array(GenNum))) | SHAInet::TrainingData, training_type : Symbol | String = :sgdm, cost_function : Symbol | String | CostFunction = :mse, epochs : Int32 = 1, error_threshold : Float64 = 0.00000001, mini_batch_size : Int32 = 1, log_each : Int32 = 1, show_slice : Bool = false, autosave : NamedTuple(freq: Int32, path: String) | Nil = nil)

  This method is kept for matching syntax of previous versions.

• #train_es(data : Array(Array(Array(GenNum))) | SHAInet::TrainingData, pool_size : Int32, learning_rate : Float64, sigma : Float64, cost_function : Symbol | String | CostFunction = :c_ent, epochs : Int32 = 1, mini_batch_size : Int32 = 1, error_threshold : Float64 = 0.0, log_each : Int32 = 1, show_slice : Bool = false, autosave : NamedTuple(freq: Int32, path: String) | Nil = nil)

  Use evolutionary strategies for network optimization instread of gradient based approach ameba:disable Metrics/CyclomaticComplexity

• #train_es_exp(data : Array(Array(Array(GenNum))) | SHAInet::TrainingData, pool_size : Int32, learning_rate : Float64, sigma : Float64, cost_function : Symbol | String | CostFunction = :c_ent, epochs : Int32 = 1, mini_batch_size : Int32 = 1, error_threshold : Float64 = 0.0, log_each : Int32 = 1, show_slice : Bool = false, autosave : NamedTuple(freq: Int32, path: String) | Nil = nil)
• #update_biases(learn_type : Symbol | String, batch : Bool = false)

  Update biases based on the learning type chosen

• #update_mse

  Calculate MSE from the error signal of the output layer

• #update_weights(learn_type : Symbol | String, batch : Bool = false)

  Update weights based on the learning type chosen

• #validate_values(array : Array(Float64), location : String)
• #verify_data(data : Array(Array(Array(GenNum))))
• #verify_net_before_train
• #w_gradient : Array(Float64)

Constructor Detail

Instance Method Detail